<p>Traditional feature extraction methods often overlook the structural characteristics of noisy data in multi-view scenarios. In this paper, we argue that noisy data in a single view can be categorized into two types in a multi-view context: true-noise data, which negatively impacts downstream tasks, and pseudo-noise data, which can be enhanced by complementary information from other views and thus exhibits higher discriminability after dimensionality reduction. Based on the varying influence of true-noise data and pseudo-noise data on feature extraction across different datasets, we propose two robust feature extraction methods. Both methods are inspired by Robust Principal Component Analysis to effectively separate high-magnitude, sparse noisy data from a single view or all views. Additionally, by incorporating nonlinear shared embedding terms and regularization constraints, the methods aim to preserve pseudo-noise data features of samples during multi-view feature extraction. Extensive numerical experiments on both clean and noisy datasets demonstrate the effectiveness and robustness of the proposed approaches.</p>